Conditioned Quantum Diffusion Models

• Conditioned Quantum Diffusion Models (CQDDs) are generative models that integrate quantum constraints with diffusion dynamics to produce valid quantum states or circuits.
• They leverage structure-preserving mirror maps and advanced conditioning mechanisms—such as ancilla encoding, label qubits, and cross-attention—to steer the generative process.
• CQDDs demonstrate robust performance across tasks like quantum state synthesis, few-shot learning, and circuit generation, yielding high-fidelity, controllable outputs.

Conditioned Quantum Diffusion Models (CQDDs) are a class of generative models that integrate quantum-theoretic structure with diffusion-based sampling, specifically targeting distributions over quantum states or circuits, and augmenting them with conditioning mechanisms. These models leverage quantum-inspired architectures, parameterized quantum circuits, or hybrid quantum–classical neural networks to construct diffusion processes whose forward and reverse dynamics respect quantum constraints and can be steered towards desired conditional distributions via explicit label, prompt, or ancillary encoding.

1. Theoretical Foundations of Conditioned Quantum Diffusion

CQDDs generalize classical denoising diffusion probabilistic models (DDPMs) to quantum state spaces or tensor representations relevant for quantum circuit synthesis and quantum machine learning (Zhu et al., 2024, &&&1&&&, Barta et al., 27 May 2025). Let $\mathcal{C}^+_n$ denote the set of $n \times n$ complex, Hermitian, positive semidefinite matrices with unit trace—i.e., the density matrices representing mixed quantum states. Standard diffusion processes perform forward noising and reverse denoising using stochastic differential equations (SDEs) or discrete Markov chains, typically over unconstrained Euclidean spaces.

Quantum generative tasks pose the challenge that the target distributions are either density matrices with strict positivity, Hermiticity, and trace conditions, or structured circuit tensors that encode gate types and continuous rotations. CQDDs incorporate these quantum constraints by embedding the quantum state or circuit representation into a dual variable via a structure-preserving mirror map (notably the gradient of the negative von Neumann entropy $\phi(X) = \operatorname{tr}[X \log X]$, with mirror map $\Phi_0(X) = I + \log X$), allowing the forward SDE to run over an unconstrained Euclidean space, and reconstructing the quantum object via the inverse map upon sampling (Zhu et al., 2024). This approach hard-wires physical constraints directly into the generative process, ensuring admissibility of sampled quantum states even for unseen data.

In quantum circuit synthesis and latent variable models, the forward diffusion operates on circuit tensors $x_0 \in \mathbb{R}^{Q \times T \times D}$ or classical vectors, adding Gaussian noise at each timestep according to a noise schedule, while the reverse process is parameterized by either neural networks or quantum circuits, possibly conditioned on text prompts, class labels, or ancillary qubits (Fürrutter et al., 2023, Barta et al., 27 May 2025, Cacioppo et al., 2023, Quinn et al., 22 Sep 2025).

2. Conditioning Mechanisms

CQDDs admit a range of conditioning strategies, enabling the model to generate quantum states or circuits correlated with explicit target distributions, physical properties, or performance metrics:

• Ancilla-based Quantum Conditioning: Conditioning information is encoded as an ancillary quantum register. For multi-class state generation, ancilla qubits prepared in RX-rotation states $$\sigma(c=\mu) = (RX(\mu)|0\rangle \langle 0| RX(\mu)^\dagger)^{\otimes n_a}$$ embed the class label as a continuous parameter, injected at each denoising step in the reverse circuit (Quinn et al., 22 Sep 2025). The joint system+ancilla unitary is followed by a projective measurement, yielding class-specific output states.
• Label Quibit Initialization: For quantum diffusion applied to few-shot learning, a label qubit is angle-encoded and processed together with data qubits using strongly entangling layers in the PQC, steering denoising towards the desired class (Wang et al., 2024). The label qubit mechanism supports conditional generation, conditional denoising inference, and label-guided noise-addition strategies.
• Cross-Attention and Embedding in Classical Neural Networks: In circuit or image-based CQDDs, textual prompts, class embeddings, or target properties are encoded via CLIP or learned MLPs and injected into U-Net denoisers via cross-attention layers or additive embedding, supporting classifier-free guidance. During training, labels are randomly dropped to regularize the network for both conditional and unconditional sampling (Fürrutter et al., 2023, Barta et al., 27 May 2025, Mauro et al., 23 Dec 2025).

A summary table highlights key conditioning modes:

Conditioning Mechanism	Quantum State Models (Quinn et al., 22 Sep 2025, Wang et al., 2024)	Circuit/Image Models (Barta et al., 27 May 2025, Mauro et al., 23 Dec 2025)
Ancilla RX-Rotation	Yes	No
Label Qubit (Angle/State)	Yes	No
CLIP/Text Embedding	No	Yes
Cross-Attention in U-Net	No	Yes
Classifier-Free Guidance	Yes (score-mix)	Yes (CFG in denoising)

All conditioning methods enable single-model multi-class generation, eliminating the need to train separate models for each class or target property (Quinn et al., 22 Sep 2025).

3. Forward and Reverse Diffusion Processes

The canonical workflow in CQDDs comprises:

1. Forward Diffusion (Noising):
   • For quantum density matrices, mirror diffusion transforms $X \in \mathcal{C}^+_n$ to an unconstrained dual variable $Y = \Phi(X)$, then evolves $Y$ by a variance-preserving forward SDE: $$dy_t = -\frac{1}{2} \beta(t) y_t dt + \sqrt{\beta(t)} dw_t$$ (Zhu et al., 2024).
   • For circuit tensors/images, elementwise Gaussian noise is added: $$x_t = \sqrt{\bar{\alpha}_t} x_0 + \sqrt{1-\bar{\alpha}_t} \epsilon$$, $\epsilon \sim \mathcal{N}(0,I)$ (Fürrutter et al., 2023, Barta et al., 27 May 2025, Mauro et al., 23 Dec 2025).
2. Reverse Diffusion (Denoising):
   • The reverse SDE (quantum states): $$dy_t = [+ \frac{1}{2} \beta(t) y_t + \beta(t) \nabla_y \log p_t(y) ] dt + \sqrt{\beta(t)} d\tilde{w}_t$$, with the score replaced by a neural approximation or quantum circuit function (Zhu et al., 2024, Quinn et al., 22 Sep 2025).
   • In PQC-based or hybrid models, the reverse kernel takes a Gaussian form parameterized by an $\epsilon$-predictor, neural or quantum: $$p_\theta(x_{t-1}|x_t,c) = \mathcal{N}(x_{t-1}; \mu_\theta(x_t,t,c), \beta_t I)$$, with $\mu_\theta$ constructed from score matching (Fürrutter et al., 2023, Barta et al., 27 May 2025, Mauro et al., 23 Dec 2025).
   • Structure recovery: in quantum-state models, invert the mirror map and enforce $\operatorname{tr} X=1$ for final sample validity (Zhu et al., 2024).

Sampling is often guided by classifier-free guidance, combining the unconditional and conditional scores as $$s_\mathrm{guided}(y,t;c,\omega) = s_\theta(y,t,\varnothing) + \omega[s_\theta(y,t,c) - s_\theta(y,t,\varnothing)]$$ (Zhu et al., 2024, Fürrutter et al., 2023, Barta et al., 27 May 2025).

4. Architectures, Algorithms, and Training Objectives

CQDDs span classical, hybrid, and fully quantum architectures:

• Score-Matching Networks (Density Matrices): An MLP with 8 residual blocks, time and label embeddings via SiLU-activated fully-connected layers, group normalization, and a final out-module (Zhu et al., 2024).
• Quantum Circuits (State or Latent Generation): PQC ansätze of $L$ layers: single-qubit rotations $R_x$, $R_y$, $R_z$, and entangling gates (CZ or CNOT) in brickwork or ring topology; ancilla or label qubits entangled and measured (Quinn et al., 22 Sep 2025, Cacioppo et al., 2023, Wang et al., 2024).
• U-Net with Cross-Attention (Circuit/Image Generation): Residual CNN blocks augmented with cross-attention on conditioning vector or text prompt, positional embeddings, and classifier-free guidance (Fürrutter et al., 2023, Barta et al., 27 May 2025, Mauro et al., 23 Dec 2025).
• Hybrid Quantum–Classical Networks (QCU-Net): Classical U-Net encoder/decoder with inserted quanvolutional and quantum ResNet blocks, extracting quantum-enhanced features from spatial patches via 12-qubit VQCs (Mauro et al., 23 Dec 2025).

Training employs denoising score-matching loss or infidelity objectives:

• Classical/Hybrid: $$L(\theta) = \mathbb{E}[ \|\epsilon - \epsilon_\theta(x_t,t,c)\|^2 ]$$ (Zhu et al., 2024, Fürrutter et al., 2023).
• Quantum Infidelity: $$L_t(\theta_t) = 1 - \mathbb{E}_{x_t,x_{t-1}}[ F(P(\theta_t)|x_t\rangle, |x_{t-1}\rangle) ]$$ (Cacioppo et al., 2023).

Quantum models optimize via ADAM and backpropagation on circuit output amplitudes, or parameter-shift rules on hardware.

5. Demonstrated Applications and Quantitative Performance

CQDDs have been validated across generative quantum state modeling, circuit synthesis, Earth observation imagery, and quantum few-shot learning:

• Quantum State Generation: Structure-preserving generation of 16$\times$16 (4-qubit) density matrices in three entanglement classes; all metrics (sliced-Wasserstein, energy-MMD, negativity Wasserstein) at $10^{-3}$–$10^{-4}$; convex combination of labels yields linear interpolation of entanglement negativity (Zhu et al., 2024).
• Quantum Few-Shot Learning: Label-conditioned CQDDs (via LGGI, LGNAI, LGDI) achieve 71.9–99.2% accuracy (vs. QNN baseline 57.4%), with large gains (20–30 pp) in few-shot classification tasks; zero-shot accuracy 60–80% demonstrates conditional prior generalization (Wang et al., 2024).
• Parameterized Circuit Synthesis: U-Net CQDDs generate PQCs for GHZ state preparation (95%+ samples $F>0.99$ on 3 qubits, 30% on 4-qubits, guidance scale $s$), QML classification (accuracy distribution matches prompt), and circuit diversity above 70% (Barta et al., 27 May 2025). Text and unitary-compilation conditioning supports efficient circuit search (Fürrutter et al., 2023).
• Quantum-Enhanced Earth Observation: QCU-Net achieves FID=2.57 and KID=$(0.8\pm0.3)\times10^{-3}$, outperforming classical U-Net baselines; semantic classification accuracy raised to 83% vs. 59% (classical) (Mauro et al., 23 Dec 2025).
• Full Quantum Conditioning: Ancilla-conditioned CQDDs yield order-of-magnitude reductions in test loss ($1.4$–$7.9$% vs 84% unconditioned) for single- and multi-qubit state distributions (Quinn et al., 22 Sep 2025, Cacioppo et al., 2023).

6. Limitations and Prospective Directions

CQDD scalability faces acute challenges:

• Quantum State Dimension Blowup: Modeling quantum states for $N$ qubits requires networks over a $2^{2N}$-dimensional manifold, quickly exceeding classical resource bounds beyond $N=4$ (Zhu et al., 2024).
• Sampling Complexity: Reverse SDEs require hundreds to thousands of steps; probability-flow ODEs or learned sampler acceleration are active research areas (Zhu et al., 2024).
• Quantum Hardware Constraints: Limited qubit counts, gate errors, and decoherence restrict circuit depth and label register width; non-selective measurements are necessary to avoid exponential sample loss (Cacioppo et al., 2023).
• Architectural Specialization: Integration of tensor networks, block-diagonal, or complex-valued networks could reduce sample complexity and meet quantum symmetries (Zhu et al., 2024).
• Alternate Mirror Maps and Objectives: Use of other entropy functionals (Tsallis, Bregman divergences) or likelihood-based score-matching could improve fidelity and geometric efficiency (Zhu et al., 2024).

A plausible implication is that hierarchical or latent CQDDs, hardware-parallel quantum modules, and multi-modal conditioning will be critical for pushing CQDD generative power to true large-scale quantum data and practical quantum machine learning regimes.

7. Context and Significance

CQDDs consolidate quantum theory and state-of-the-art generative modeling, providing models with explicit physics priors, multi-class conditional controllability, and generalization across circuit structures and gate alphabets. By embedding structure-preserving maps and leveraging both quantum circuit architectures and advanced neural attention mechanisms, CQDDs have demonstrated high-fidelity quantum state and circuit synthesis, robust few-shot learning, and empirical gains in quantum-enhanced remote sensing imagery. The unified framework for conditioning enables parameter sharing, significant error reduction, and fine-grained generative control, reaffirming CQDDs as central to the intersection of quantum simulation, generative modeling, and quantum machine learning (Zhu et al., 2024, Fürrutter et al., 2023, Barta et al., 27 May 2025, Wang et al., 2024, Quinn et al., 22 Sep 2025, Mauro et al., 23 Dec 2025, Cacioppo et al., 2023).